Network-based drug repurposing for psychiatric disorders using single-cell genomics

Summary

Neuropsychiatric disorders lack effective treatments due to a limited understanding of the underlying cellular and molecular mechanisms. To address this, we integrated population-scale single-cell genomics data and analyzed 23 cell-type-level gene regulatory networks across schizophrenia, bipolar disorder, and autism. Our analysis revealed potential druggable transcription factors co-regulating known risk genes that converge into cell-type-specific co-regulated modules. We applied graph neural networks on those modules to prioritize novel risk genes and leveraged them in a network-based drug repurposing framework to identify 220 drug molecules with the potential for targeting specific cell types. We found evidence for 37 of these drugs in reversing disorder-associated transcriptional phenotypes. Additionally, we discovered 335 drug-cell quantitative trait loci (eQTLs), revealing genetic variation’s influence on drug target expression at the cell-type level. Our results provide a single-cell network medicine resource that provides potential mechanistic insights for advancing treatment options for neuropsychiatric disorders.

Graphical abstract

Highlights

• •Cell-type-specific gene regulators in psychiatric disorders are identified
• •Graph neural networks identified novel risk genes and drug candidates
• •Evidence of therapeutic effect for 37 drugs
• •Discovered 335 cell-type-specific eQTLs linking genetic variation to drug targets

Using single-cell genomics and advanced network analysis, Gupta et al. mapped gene networks in key brain cell types across schizophrenia, bipolar disorder, and autism. This innovative approach uncovered druggable transcription factors and over 200 repurposable drug candidates, pointing to promising new targeted treatments for these major psychiatric disorders.

Introduction

Psychiatric disorders, including schizophrenia (SCZ), bipolar disorder (BPD), and autism spectrum disorder (ASD), are complex conditions lacking effective treatments due to complex gene regulation and cellular heterogeneity.¹ Single-cell multiomics offer a promising avenue for analyzing dysregulated epigenomic and transcriptional states and identifying specific cell types affected by the disease.^2,3,4,5 Analysis of cell-type gene regulatory networks (GRNs) and dysregulated pathways that underpin disease can lead to identifying novel therapeutic targets for specific cell types.^6,7 This resolution enables drug-repurposing efforts to focus on modulating abnormalities in gene regulation within specific cell types, thereby facilitating the discovery of effective treatments that are tailored to the underlying etiology of each disease. However, post-genomic drug-repurposing efforts for cell types perturbed in neuropsychiatric disorders remain poorly characterized, mainly due to the lack of uniformly processed datasets that allow investigation across disorders using control and disease groups.

This gap was recently filled by the PsychENCODE phase II (PEC2) study.⁸ The phase II efforts of the consortium generated single-cell sequencing data from prefrontal cortex of adult human brains with multiple neuropsychiatric disorders. The collection provides uniformly processed single-nucleus RNA sequencing (snRNA-seq), single-nucleus assay for transposase-accessible chromatin sequencing (snATAC-seq), and snMultiome (paired snRNA-seq and snATAC-seq). The dataset comprises >2.8M nuclei across 28 distinct brain cell types from 388 individual brains. The dataset accompanies 24 cell-type GRNs inferred using data from healthy donors. The cell-type GRNs were found to be significantly different between cell types, with extensive rewiring and differential usage of enhancers and promoters by the same transcription factors (TFs) and co-regulation patterns of disease-risk genes. These previous observations invite a more thorough, cross-disorder GRN comparison between cell types to highlight key differences in the regulatory dynamics of psychiatric disorders, as these lines of investigation had not been explored in the original PEC2 work. The PEC2 dataset offers valuable single-cell multiomics data that can support the evaluation of network-based drug repurposing in the context of cell types relevant to neuropsychiatric disorders.

GRNs offer a powerful means of linking promoters and enhancers with interactions between genes and their regulators (such as TFs). TFs effectively act as molecular “switches” that control the transcriptional output of cells and play important roles in disease phenotypes,⁹ thereby making them promising therapeutic targets.⁹ Recent advances in ligand engineering have challenged the notion of TFs as being “undruggable,” a perspective that can mainly be attributed to their structural and biophysical complexity.^9,10,11 TFs can be interesting targets for drugs designed to restore transcriptional phenotypes¹² or selectively repress or activate disease regulons (gene sets under similar regulatory control)^13,14,15 or when combining two or more drugs with different mechanisms to boost clinical efficacy.^16,17 Several TFs have been successfully targeted in disease treatment. For instance, tamoxifen, a selective estrogen receptor modulator (SERM), acts as an estrogen receptor antagonist in breast cancer by blocking estrogen-mediated signaling, while enzalutamide targets the androgen receptor in prostate cancer. Pioglitazone and rosiglitazone activate PPARγ, improving insulin sensitivity in type 2 diabetes. While TF-targeting drugs are established in oncology and metabolic diseases, psychiatric applications are still emerging.

While there are several approaches for using genomic data for drug repurposing,¹⁸ network-medicine-based strategies have been widely adopted for brain-related disorders such as Alzheimer’s and SCZ^19,20,21,22 as well as for other human diseases, including respiratory and cardiovascular conditions.^23,24,25 The strategy aims to test the network proximity between disease gene sets and drug targets in biological networks such as protein-protein interactions, thus increasing the likelihood that they affect the disease through multiple network pathways. Thus, the availability of population-scale single-cell sequencing data can enhance network-medicine-based strategies by revealing key dysregulated mechanisms at the cell-type level. Furthermore, an integrated analysis of single-cell multiomics, genetic variants, and drug targets can be beneficial for applications in personalized medicine.

However, the identification of effective drug targets for brain disorders remains hindered by limited knowledge of disease-risk genes and their cell-type specificity. Many disease genes may be co-regulated by similar regulatory mechanisms,^26,27,28 suggesting that cell-type GRNs contain the information necessary to bridge the gap between known risk genes and novel candidate genes. Graph learning approaches, which leverage network information to identify similar nodes and prioritize edges, have successfully inferred disease genes in various networks, e.g., protein-protein interactions.^29,30,31,32 However, the application of graph learning to cell-type GRNs for inferring novel candidate disorder genes, especially at the cell-type level, remains largely underexplored. Novel risk genes with cell-type specificity can help identify relevant drug targets.

While previous studies have explored drug repurposing for psychiatric disorders, most were based on bulk brain tissue and lacked cell-type-specific resolution. For instance, regulatory network approaches identified drug candidates for SCZ and ASD,^19,33 and others integrated genome-wide association study (GWAS) data with drug-target databases to highlight targets like interleukin 6R (IL-6R) for SCZ and BPD³⁴ or used machine learning to prioritize SNPs linked to depression.³⁵ We still lack a systematic survey of repurposable drug candidates for cell types implicated in neuropsychiatric disorders. Single-cell transcriptomics enables the identification of cell-type-specific transcriptional changes that are often masked in bulk tissue analyses, allowing for more precise mapping of disease mechanisms and targeted drug repurposing.⁷ To address this gap, our study explores a single-cell network-medicine-based approach to drug repurposing for three neuropsychiatric disorders: ASD, SCZ, and BPD. Our study is divided into four main parts. First, we infer cross-disorder cell-type GRNs by integrating snRNA-seq data and snATAC-seq single-cell data from a population of 140 diagnosed individuals and 107 age-matched controls in the PEC2 cohort. We then survey network properties that characterize these GRNs and identify differential TF regulons enriched with drug targets. Second, we utilize the GRNs to train graph neural networks (GNNs) to predict cell-type-specific disorder-risk genes. Third, we leverage these prediction models to identify and prioritize drugs that could potentially restore the transcriptional phenotype of neuropsychiatric disorders to a healthy state. Finally, we identify genetic variants that explain intrinsic expression variation of repurposable drug targets in an attempt to find drug-associated cell-type expression quantitative trait loci (eQTLs). Our study allowed us to explore the potential genetic associations with approved drugs in complex brain diseases and traits, thereby contributing to the foundation of precision medicine.

Results

Cell-type gene regulatory network construction for drug repurposing in psychiatric disorders

We leveraged the recently published PsychENCODE2 dataset for network-based drug repurposing to study psychiatric disorders (Figure S1). This cross-cohort, uniformly processed dataset enables us to predict high-quality cell-type GRNs across three psychiatric disorders: ASD, SCZ, and BPD (Figure 1A). We utilized >1.64 million cells from 247 donors and applied a GRN inference pipeline to comprehensively analyze dysregulated transcriptional pathways, drug repurposing, and pharmacogenomics across 23 cell types and the three psychiatric disorders (Figures 1 and S2A). We identified rewired regulatory genes that are influenced by specific drug treatments and patterns of co-regulated disorder-risk genes and their convergence into modules (Figure 1B). We also utilized a machine learning framework that predicted 249 novel cell-type-disorder-risk genes (Figure 1C). Furthermore, using the cell-type-disorder GRNs and predicted disorder genes, we tested a library of >800 drug molecules for their proximities to each disorder within each cell type. Our framework nominated 220 drugs, 38 of which have existing evidence for the potential to modulate the transcriptional phenotype of the disorders (Figure 1D). This approach enabled us to link prioritized drugs to individual donors based on their genotypes, allowing for the identification of drug-cell eQTLs with applications in pharmacogenomics and precision medicine (Figure 1E). We have made our analyses available on a web application hosted at https://daifengwanglab.shinyapps.io/psyDGN/.

Figure 1.

Network-based drug repurposing workflow

(A) The PEC2 dataset was utilized to build gene regulatory networks for three psychiatric disorders across 23 cell types of the human brain.

(B) The GRNs link TFs to target genes via proximal interactions (from snRNA-seq) or distal interactions (from snATAC-seq) under both control (healthy) and disorder conditions from age-matched donor groups. The GRNs are then analyzed to identify druggable regulons and target gene modules that show strong statistical enrichment of known risk genes.

(C) We trained several variants of supervised graph neural networks (GNNs) to predict and rank novel risk genes in a cell-type-specific manner.

(D) A network-proximity-based procedure was applied to probe drugs with targets in close vicinity to risk genes predicted by the GNNs.

(E) Identification of eQTLs that affect drug target gene expression.

Druggable cell-type transcription factors for psychiatric disorders

We integrated snATAC-seq and snRNA-seq data to predict cell-type GRNs in ASD, SCZ, and BPD, as well as in their respective control (healthy) individuals (see STAR Methods). Briefly, we linked TFs to their potential target genes based on the strength of co-expression observed in the snRNA-seq data, thereby identifying proximal links. Additionally, we connected TFs to interacting enhancers and promoters in the snATAC-seq data to derive potential distal links (Figure 2A). Overall, we predicted GRNs across eight major cell types and 23 subclasses (see STAR Methods). Overall, we inferred 69 GRNs, which include an average 142 TFs and 5,000 target genes linked via an average of 5,840 proximal and 800 distal links (Figures S2B and S2C). We found a relative overenrichment of distally regulated drug targets in inhibitory neuronal cells of ASD individuals compared to the other two disorders (Figure S2D), which makes these targets compelling to studies that seek to disrupt disease-associated enhancers (PMID: 35132191). Cross-cohort analysis showed strong concordance between cell-type GRNs (Figure S3), suggesting the GRNs are insensitive to sampling bias. Furthermore, among all the predicted TF-target links, ∼62% were also present in an independent chromatin immunoprecipitation sequencing (ChIP-seq) assay,³⁶ suggesting the presence of TF binding sites on the regulatory regions of their predicted targets (Figure S3B).

Figure 2.

Druggable cell-type transcription factors

(A) Pipeline to integrate snRNA-seq and snATAC-seq data to predict TF → target gene links.

(B) An excitatory neuron subnetwork plot showing differential usage of target genes by TFs in the different psychiatric disorders.

(C) TFs (x axis) that act as hubs specifically in cell-type (x axis, bottom) and disorder (x axis, top) GRNs. TF names are hidden for ease in visualization.

(D) Circular heatmaps showing the distribution of TF disorder influence (DI) scores across cell types and disorders. The outermost, middle, and inner rings show distributions in bipolar disorder (BPD), schizophrenia (SCZ), and autism (ASD), respectively. The names of cell types along the x axis and TFs along the y axis are hidden for ease in visualization (the full version is shown in Figure S5). The cell types of the heatmap are colored along a blue-to-red gradient representing gain or loss in influence, respectively. Only the outer ring of the heatmap is clustered; the inner two rings remain unclustered to preserve order.

(E and F) TFs with the largest changes in DI score are shown on the y axis along with their DI score (right bar plot) and log of change in expression in disorder with respect to control (left bar plot). The heatmaps show a change in the expression of TFs in CNS cells treated with drug compounds (x axis; names hidden) represented along a blue-to-red gradient indicating upregulation or downregulation, respectively.

We found extensive rewiring of TF regulons (the set of predicted target genes for a TF) between disorders (Figure 2B). For instance, the FOS TF, a critical regulator of neural activity, exhibits differential usage of enhancers and promoters across different disorders (Figure 2B). We generated a list of cell-type- and disorder-specific hubs (TFs with a disproportionately large number of targets) and bottlenecks (TFs that act as bridges) (Figures 2C and S4A–S4C; Table S1). Among all hubs, 22% were common between the three disorders, and 20% were unique to SCZ and ASD (Figure S4C). The bottleneck TFs are those with a central role in connecting different parts of the network, measured by “betweenness centrality.” This score indicates which TFs are most influential in regulating other genes. To understand how these TFs change in different disorders, we compared their centrality scores in healthy (control) versus disorder networks, creating a “disorder influence” (DI) score. This DI score shows how much influence a TF gains or loses in each disorder (see STAR Methods). The DI score revealed distinct cell-type-specific patterns in our GRNs (Figure 2D). For instance, we found excitatory neurons to be the cell type with the most elevated DI scores in ASD and BPD (Figure S4D), which aligns with previous observations.³⁷ Interestingly, our analysis showed that microglia exhibit the highest DI scores across TFs in SCZ (Figure S4D), suggesting a substantial shift in regulatory dynamics within this cell type. This aligns with previous studies linking SCZ to excessive synaptic pruning and neuroinflammation,^38,39 processes in which microglia play a key role. The DI score helped us narrow down a list of the most rewired TFs for further interrogation (Table S2).

Furthermore, since TFs are not readily available as candidates for drug targets, we hypothesized that certain drug compounds may confer an indirect effect on TFs, influencing their expression activity, which can be monitored in disease cells. Indeed, we found several examples of TFs with high DI scores whose expression activity in cell lines treated with certain drug compounds was opposite the observed expression in the three disorders we studied (Figures 2E, 2F, and S4E; Table S3).

Overall, our GRNs show extensive rewiring of TF nodes with differential usage of target genes (Figure S5). Our GRNs suggest a complex and heterogeneous landscape of gene regulation in psychiatric disorders, where different cell types and TFs may contribute uniquely to disease etiology. Understanding these distinct regulatory mechanisms at the level of target genes can provide a more holistic perspective into the molecular underpinnings of these disorders and potentially inform more targeted therapeutic strategies.

Disorder-risk genes converge on co-regulated gene modules

Since most known disease genes (and drug targets) are non-TF genes, it is imperative to analyze the GRNs at the target gene level to fully capture the regulatory dynamics and therapeutic potential of GRN models. We transformed our directed cell-type GRNs to undirected co-regulatory networks, which connect target gene pairs that have high overlaps between their regulators (see STAR Methods). We then clustered these networks to derive modules of highly co-regulated genes, yielding >4,000 dense modules across cell types and disorder networks (Figure S6A; Table S4), several of them enriched with Gene Ontology (GO) terms representing biological processes (Figures S6B–S6D; Table S5).

We then assessed whether these gene modules are preserved between control and disorder networks, using normalized mutual information (NMI) to compare clustering patterns. NMI values range from 0 (dissimilar) to 1 (similar), allowing us to identify distinct cell-type-specific patterns. For instance, modules in inhibitory neurons remain relatively unchanged in ASD and SCZ but exhibit drastic changes in BPD (Figure 3A), whereas SCZ perturbs excitatory neurons the most (Figure 3A). These patterns align with known biological insights related to these disorders. We also tracked the number of modules enriched with known risk genes and found the largest number of such modules in ASD (Figure 3A), likely due to relatively well-characterized knowledge bases for ASD.

Figure 3.

Prediction and analysis of co-regulated gene modules

(A) Normalized mutual information (NMI) is used as a measure to gauge the deviation of network clustering between control and disorder networks. The heatmap shows the three disorders (x axis), the cell types (y axis), and grids filled along a gray-to-white gradient representing low (dissimilar clustering) or high (similar clustering) mutual information, respectively. The inner squares are sized according to the number of disease modules identified within each cell type × disorder network.

(B) A zoom-out of microglial modules in ASD. The heatmap at the bottom shows modules (on the x axis) and their differential expression⁴⁰ in sex, age, and whole-cortex contrasts in ASD samples (y axis). The grids of the heatmap are colored along a red-to-blue gradient indicating down- and upregulation, respectively. The columns of the heatmap are annotated for the enrichment of candidate ASD genes (from SFARI⁴¹) and GWAS-implicated loci in psychiatric disorders within the identified microglial modules (x axis). Only three modules described in the main text are labeled.

(C–E) Visualization of modules M52, M38, and M6. Each circle represents a gene colored according to the log-fold change value in the whole cortex in ASD donors versus healthy controls. Borders of known risk genes are colored cyan and drug targets are shaped as diamonds. Genes are connected by edges representing shared regulators.

Recent studies reveal shared genetic loci among these disorders, highlighting substantial overlap in their genetic underpinnings despite diagnostic and phenotypic complexity.^42,43 Building on this, we further explored ASD modules to examine whether these modules are enriched for genetic variants (SNPs) associated with other psychiatric disorders. Using a heritability-aware enrichment analysis framework⁴⁴ and data from published GWASs, we identified several ASD microglial modules significantly enriched with SNPs linked to BPD, SCZ, and epilepsy (Figure 3B). For instance, we recovered an age-specific ASD module, M52 (Figure 3C), containing known ASD candidate genes such as IKZF1.⁴⁵ These findings suggest a convergence of genetic risk factors from multiple psychiatric disorders into common molecular modules, implying underlying shared molecular mechanisms, potentially underpinning overlapping clinical symptoms such as depression.⁴⁶ One particular module of interest is module 38 (referred to as M38), which shows enrichment of both ASD GWASs, high-confidence gene sets from the SFARI database and bipolar GWASs (Figure 3B). Genes within M38 are collectively downregulated in whole-cortex samples of ASD but remain relatively unchanged in age or sex (Figure 3B). MET, a promising candidate gene for ASD^47,48 and a target for a blood-brain barrier (BBB)-permeative drug, is a member of M38 (Figure 3D). Another interesting module with an incidence of hits across the three GWASs is M6, with downregulated genes in whole-cortex ASD (Figure 3B) and membership of candidate ASD genes such as GRIK5⁴⁵ and CDH11⁴⁹ (Figure 3E).

Collectively, our module analysis suggests that certain candidate risk genes are co-regulated in cell types. We find evidence of psychiatric disorders converging on a few of these modules, perhaps relevant to clinical phenotypes common across these disorders (e.g., depression) and possibly influenced by drug compounds with well-defined modes of action. This set of modules therefore provides a useful resource for more systematically analyzing drugs that may be repurposed to target specific cell clusters.

Graph neural networks predict novel cell-type-disorder genes

Most reported disease-risk genes in the literature lack information on cell-type specificity, which is a prerequisite for our main goal of cell-type drug repurposing. Our analysis of network modules provides insight into genes that co-regulate with known candidate risk genes in certain cell types. Typically, a drug repurposing pipeline proceeds by analyzing a single disease module (identified based on enrichment of known risk genes). However, such analyses rely heavily on the assumption that all genes within a disease module are equally equitable risk genes while leaving out several other risk genes that are distributed across several modules and vary based on the selected clustering algorithm. We addressed this limitation by using supervised graph convolutional networks (GCNs) training on cell-type-specific disorder GRNs, where genes are represented as nodes and known risk genes, identified from public databases, are labeled. This allowed us to leverage both known risk genes and cell-type networks as training data. Briefly, the GCN works as follows: a combination of graph convolution and fully connected layers is used to generate embeddings for each gene and classify genes as risk or non-risk genes (Figure 4A; Table S6). The models were trained using 5-fold cross-validation. After validation, the average of all GCN disease scores is used for subsequent analyses.

Figure 4.

Graph learning model prioritizes existing candidate disease genes and novel risk genes in ASD, BPD, and SCZ

(A) Schematic for gene classification using the graph learning model, i.e., graph convolutional network, featuring risk (red), non-risk (blue), and unknown (gray) genes.

(B) Performance of the model with specified targets. Circle sizes scale from radii of 0–1 based on AUPRC for baseline (solid line), model 1 (disease genes only, dashed line), and model 2 (disease and module genes, filled circle).

(C) Number of targets of blood-brain barrier (BBB) drugs found in the top decile of prioritized genes by model 1 (dashed line) and model 2 (filled bar) from 0 to 20.

(D) Histogram of the number of SFARI candidate category 2 genes found in each decile of ASD-predicting models. The top decile is annotated with the top four prioritized candidate genes. The model score (right axis) is also plotted against the prediction percentile (gray line).

(E) Top three prioritized genes from each disease and their respective prioritizations within each cell type. Missing entries were not included in the corresponding cell-type GRN. Cell types are categorized by disease, and all scores are shaded based on per-model normalized risk scores. Darker squares indicate a higher model score.

(F) Heatmap shows linkage disequilibrium-aware enrichment of GWAS traits (y axis) within the respective first-decile predictions across cell types (x axis). The grids are colored along a black gradient that is set to be proportional to the enrichment p values, with darker colors indicating greater significance.

For (B)–(E), results for additional cell types may be found in Figures S7A–S7E.

We trained two models for each cell type based on different groups of risk genes. Model 1 was trained only on known disease genes, while model 2 incorporated genes from the modules discussed above. More details on the exact derivation of the gene lists can be found in the STAR Methods. We see that the area under the precision recall curve (AUPRC) of model 2 is higher than that of model 1 in 21 of 24 cases for major cell types (Figure 4B). Subclasses show a similar pattern, with 26 of 51 cell-type-disease combinations exhibiting greater AUPRC for model 2 and 17 showing equal performance (Figure S7A). To verify that this increased accuracy on modules does not come at the expense of biological relevance, we further examined the frequency of high prioritization for known targets of drugs that can cross the BBB. Specifically, we compared the counts of BBB drug targets within the top decile of prioritized genes across models 1 and 2 (Figure 4B). In general, the performances of models 1 and 2 were near equivalent. Model 2 was favored in 11 cases, with 12 cases favoring model 1, out of a total of 24. This trend continued with subclasses, where 26 of 51 cases showed equal performance between models 1 and 2 (Figure S7B). Overall, model 2 achieved high AUPRC and prioritized similar BBB drug targets compared to model 1. Going forward, model 2 was used for our analyses.

We have shown that incorporating genes from GRN modules improved the performance of our model, but it is also necessary to demonstrate the model’s reliability in predicting novel genes. We intentionally withheld SFARI candidate genes (category 2) from our ASD model to use for validation. When plotting candidate genes’ prioritizations, we observe a heavy skew toward the top decile (Figure 4D). This skew is most pronounced in microglia and oligodendrocyte progenitor cells (OPCs) but is also seen in all major cell types aside from astrocytes and is weak in excitatory and inhibitory cell types except for Pax6 and Sncg (Figures S7C and S7D). Our GCN is, therefore, able to effectively classify unseen risk genes^50,51,52 with a high degree of reliability.

Finally, we predicted novel genes using the trained models. Averaging gene prioritization across all GRNs allows us to rank-order relevant genes and examine them for connections to ASD, BPD, and SCZ. In particular, we can take genes in the top decile of predictions for each cell type and disease and then pick only those genes that are present in all diseases (repeated for each cell type). This provided a list of 249 genes that were not positively labeled for one or more of the surveyed diseases (Figures 4E and S7E; Table S6). For instance, CTNND2, a gene that encodes an adhesive-junction-associated protein of the armadillo/beta-catenin superfamily, is a listed category 2 gene in the SFARI gene database and thus was not used as a part of the training dataset. Yet, our model prioritizes CTNND2 for ASD. There is also more recent evidence that suggests the involvement of this gene in autism.^53,54 Other examples include STXBP5, which has been noted as a seizure-risk gene⁵⁵ with more recent evidence in ASD cases.⁵⁶ SLC6A15 is additionally linked to major depressive disorder⁵⁷ and has been prioritized in ASD and BPD within overlapping gene sets for ASD, BPD, and SCZ.⁵⁸ We also checked the enrichment of genetic variants associated with the disorders in their respective GWASs within the top predictions. Interestingly, we found a greater enrichment of SNPs within non-neuronal cell types for ASD and BPD (Figure 4F). Notably, BPD also showed significant enrichment in predictions for chandelier cells. For additional validation, we repeated our ASD candidate gene analysis with SCZ and BPD, also comparing with significant SNPs from multimarker analysis of genomic annotation (MAGMA) (Figures S8 and S9).

Network-based drug repurposing

Network-based drug repurposing is an approach using gene networks to bring existing molecules to new indications. This approach has been previously shown to be effective using cell-type-naive protein networks and cross-validated using EHR data.²² The unique dataset we have at hand presents an excellent opportunity to test the feasibility of drug repurposing to target specific cell clusters in psychiatric disorders. To do this, first, we assembled a BBB-permeative drug-target network from public sources. This drug-target network has 802 drug molecules targeting 1,192 genes (Figure S10; see STAR Methods; Table S7). We observed an enrichment of known risk genes within this network, several of which are targeted by many drugs, especially GABA receptors (Figure S10). We then leveraged the cell-type co-regulatory networks described above, together with predicted risk genes from the GNN model, for network-based drug repurposing using this drug-target network.

We calculated the minimum number of edges that need to be traversed to go from a drug-target set to a risk-gene set (as predicted in our models) within a given cell-type co-regulatory network. The hypothesis, as illustrated in Figure 5A, is that if the average path length between a drug target and a disease gene is short and non-random, the drug is likely to have a direct or close indirect effect on the disease pathway. We applied this procedure to all cell-type networks we built across the three disorders. We found a total of 220 molecules that were significantly proximal (permutation-based tests, n = 100; Z score ≤ −3; Table S8) to the predicted disorder-risk genes. We found the largest distribution of repurposable drugs for SCZ and the smallest for ASD, which is counterintuitive given the stronger community confidence in known ASD-risk genes in the literature. Overall, we found more repurposable drugs for deeper-layer excitatory neurons (L6/IT, L6/CT, and L6/IT/Car3), a few inhibitory neurons such as SST and LAMP5, and microglia (Figures S11A and S11B). Collectively, a large fraction of the drugs that we found repurposable are neurotransmitter receptor modulators and reuptake inhibitors (Figure 5B). For instance, the largest category of drugs belongs to serotonin reuptake inhibitors, which are widely used to treat depression. We accurately recovered a number of drugs, such as asenapine, an atypical antipsychotic used for the treatment of SCZ and acute mania associated with BPD. We found that asenapine has close proximity to SCZ-risk genes, mostly for neuronal cell types (Figure S11C). Another interesting example of a correctly recovered drug for SCZ is iloperidone, an antipsychotic that is already used to treat SCZ symptoms. Our data suggest that iloperidone reverses gene expression, especially in several neuronal cells (Figure S11D). Importantly, we observed that 99% and 81% of the drugs nominated by our study are retained when the GRNs are trimmed 25% and 50%, respectively. Furthermore, when changing the underlying GRN inference method, ∼55% of the drugs are retained. Overall, these observations suggest that drug repurposing predictions exhibit only modest sensitivity to the choice of underlying GRN, indicating the robustness of our approach across different network inference methods (see Figure S12).

Figure 5.

Network-based drug repurposing

(A) An illustration depicting the network-based drug repurposing approach used in this study. Network proximity between cell-type-disorder genes (predicted in Figure 4) and listed drug targets was calculated across all cell-type-disorder combinations. The observed proximity was then tested against a random background from which a Z score was derived. Repurposable drug candidates were chosen based on thresholding these Z scores.

(B) Sankey plot showing links from disorders (vertical axis 1) to cell types (vertical axis 2), drug molecules (vertical axis 3), and their mode of action (vertical axis 4) as predicted using the approach illustrated in (A).

(C) The repurposable drug candidates were tested against gene expression data in the LINCS1000 dataset to determine whether the drug mimics or reverses the expression of gene sets observed under disorder conditions in the PEC2 dataset⁸ (see STAR Methods). The heatmap shows cell types and disorders on the y axis and a collection of top drug molecules (based on Z scores) on the x axis. The grids are divided into three columns, one for each disorder, and colored along a red gradient for mimickers and blue gradient for reversers. The color gradient is proportional to the −log10(adjusted p) from a two-sample Kolmogorov-Smirnov (KS) test.

To verify whether the drugs nominated by our approach can modulate the transcriptional phenotype of cells, we used the gene expression profiles from the Integrated Network-Based Cellular Signatures (LINCS) project.⁵⁹ The LINCS L1000 dataset provides gene expression measurements from various human cell lines that have been treated with thousands of small molecules. By comparing the direction of gene expression changes induced by disorders with those caused by drugs, we identified 37 drugs in our nominated drug list that mimic or reverse the gene expression patterns caused by disorders in a cell-type-dependent manner (Figure 5C; Table S9). Specifically, mimickers are drugs that replicate the gene expression changes observed in the disorder, aligning in the same direction. In contrast, reversers are drugs that counteract these changes, shifting gene expression in the opposite direction, which suggests a potential to mitigate or correct the molecular disruptions caused by the disorder. For example, genes that are differentially expressed upon treatment with fluoxetine (a selective serotonin reuptake inhibitor [SSRI] that is used to treat depression, obsessive-complusive disorder, and BPD) show opposite expression patterns compared with those observed in chandelier cells from SCZ donors, endothelial cells from ASD donors, and L2/3 IT neurons from BPD donors (Figure 5C). Fluvoxamine, another SSRI, shows potential for targeting SCZ-perturbed Lamp5 cells and BPD-perturbed microglial cells (Figure 5C).

Drug-cell eQTLs

Pharmacogenetics is a well-developed field focused on how different patients exhibit disparate responses to therapeutic compounds and how these patient-specific differences may be driven by differences in genotype. For instance, one critical mechanism by which genotypic differences may influence differences in drug response is one in which genotypic variation drives changes in the expression of drug’s target genes. As a demonstration of this idea, a very simple example scenario may be one in which an eQTL manifests as a variant that is associated with substantially diminished expression of a gene targeted by a particular drug. Under this simple scenario, patients who harbor that variant may fail to respond to therapy, given that the drug would effectively have no target gene on which to act. Critically, these general types of mechanisms are likely to play out at the level of distinct cell types.

The catalog of BBB drug targets used in our study (combined with genotypic as well as cell-type-level expression data) provides uniquely powerful building blocks to explore pharmacogenetic phenomena of this nature in greater depth than previously possible. As different therapeutic compounds affect different sets of genes, we sought to evaluate the extent to which genotypic variation is associated with significant changes in expression for the gene sets that are targeted by different drugs. To do so, we formulated a “drug-cell eQTL” analysis by measuring the extent to which expression variation of drug targets is associated with variants throughout the genome. One initial motivation for this analysis was the observation of an up to 17% overlap between drug targets and genes in our recently reported cell-type eQTLs⁸ (Figure 6A), which indicates that a substantial proportion of druggable gene targets exhibit expression that is strongly associated with genetic variation in specific cell types.

Figure 6.

Drug-cell eQTLs

(A) Fraction of the number of cell-type cis-eQTL eGenes⁸ that are also listed targets of BBB-permeative drugs.

(B) Manhattan plot showing p values (y axis) and variants’ genomic coordinates (x axis) from the drug-cell eQTL analysis. The analysis was performed using expression values of drug targets for those drugs that were found to be repurposable in our analysis.

(C) Statistically significant drug-associated cell-type eSNPs (shaped “V”) are plotted alongside associated drugs (octagons), their target genes (rounded rectangles), and associated cell types (circles) as a network graph.

(D) Distribution of the numbers of SNPs (x axis) associated with drug-cell eQTLs (y axis).

For each drug in each cell type, we used its set of target genes to build a single composite gene expression value by simply calculating the mean log-normalized expression of a given drug’s targets within each cell type. Thus, instead of using the expression of a single gene as the phenotype, a phenotype is given as a composite gene expression score for a drug-cell-type pair. Under this framework, we identified 335 eQTLs, which are defined as significant drug-cell eQTLs (Table S10). Here, significance was determined using a Bonferroni-corrected p value threshold of 0.05, with Bonferroni correction being applied at the level of each phenotype—that is, for a given drug-cell-type combination, Bonferroni correction was applied by adjusting for ∼2 million tests, where ∼2 million is the number of variants tested in our search for eQTLs (further details are given in the STAR Methods). These eQTLs may be interpreted as cases wherein a drug-associated gene set collectively exhibits strong associations with genotypic variation. As such, they highlight cases for which responses to drug exposure may vary considerably across samples, with this variation being strongly influenced by genotype. The results from this analysis are summarized in Figure 6B.

We performed GO enrichment analysis of genes linked in the eQTLs and found that the most enriched terms for molecular function and biological processes include those related to receptor activity and cell signaling (Figure S13; Table S10). We also checked for overlaps of the variants identified in drug-cell eQTLs with those in the Psychiatric Genomics Consortium (PGC) and found a skewed distribution toward smaller p values, confirming the association of these variants with psychiatric disorders (Figure S14). To provide further insight into the genes associated with the drug-cell eQTLs found in this study, we identified the set of genes that are jointly associated with both these eQTLs and those derived from single-cell data (i.e., scQTLs) by Emani et al.⁸ We identified a set of seven genes that are common to both eQTL types (see Table 1). Notably, many of these genes are directly involved in neurotransmitter signaling, nervous system activity, or the metabolism of commonly prescribed psychiatric medications. In addition, we also cataloged common sets of genomic variants between the drug-cell eQTL and the scQTLs and reported 100 such shared variants in Table S10.

Table 1.

Genes common to both scQTLs and drug-cell eQTLs

Gene	Description	Reference
ADRA1A	this is a GPCR that is involved in the contraction of smooth muscle tissue, the release of neurotransmitters, and vasoconstriction	Kolberg et al.50
CHRM2	CHRM2 plays roles in the plasticity of synapses, neuron excitability, and regulating the release of acetylcholine	Wang et al.52
CHRM5	these receptors are involved in responses to acetylcholine within both the peripheral and the central nervous systems	Bonner et al.60
CYP2D6	CYP2D is involved in metabolizing a variety of therapeutic compounds, such as antipsychotics, beta-blockers, and antidepressants	Snider et al.61
DRD4	this GPCR plays central roles in the mesolimbic system; it is involved in dopamine signaling	Czermak et al.62
MET	c-Met (a receptor tyrosine kinase) plays critical roles in development and repair and is also involved in cell growth	Park et al.63
NFE2L2	NFE2L2 is a transcription factor that is essential for responding to cellular oxidative stress, and it aids in regulating genes responsible for detoxification	Huang et al.64

We visualized statistically significant expression single nucleotide polymorphisms (eSNPs) along with the connected drugs, target genes, and associated cell types as a network, which revealed 12 of the 23 cell types we included in our analysis, with oligodendrocytes a prominent hub (Figure 6C). Interestingly, we found the largest number of variants linked to propiomazine and promazine, two antipsychotics used to treat positive and negative symptoms of SCZ (Figure 6D). This suggests that genetic variation between individuals may influence responses to these specific drugs, potentially offering insight into variability in drug efficacy.

Interestingly, we also found other potentially repurposable drugs that are not currently used to treat psychiatric disorders. For instance, benzquinamide, a hub in our network, is an antihistamine with sedative properties. Benzquinamide’s ability to antagonize dopamine and histamine receptors suggests that it could have some potential impact on psychiatric symptoms. Histamine H1 antagonism can lead to sedation, which might be useful for patients with agitation or insomnia, which frequently accompany psychiatric disorders. However, these claims remain subject to intensive clinical trials, which is beyond the scope of our study.

Integrative prioritization of druggable targets

We next sought to integrate the lines of evidence from our analyses into a single composite score to help further prioritization of key targets. We performed a systematic integration focused on genes targeted by drugs exhibiting mimicker-reverser relationships in our study. For the 37 such drugs, we evaluated each gene target across four distinct analytical layers, including (1) differential expression status from PsychENCODE2 snRNA-seq data, (2) gene-level MAGMA analysis assessing enrichment of GWAS signals within disease-associated modules, (3) GNN predictions of candidate risk genes, and (4) associations from drug-cell eQTL analyses (see STAR Methods). We ranked genes by their total support to identify high-confidence candidates (Table S11).

Notably, the top-ranked genes were highly consistent with known biology. For instance, the top candidate, BDNF, encodes a neurotrophic factor crucial for neuronal development and plasticity⁶⁵ and is a known target of antidepressants in the DrugBank database. Similarly, DRD1 and DRD2, which encode dopamine receptors implicated in multiple neuropsychiatric disorders,^66,67 and GRIK2, a glutamate receptor gene,⁶⁸ were also prioritized by our integrative framework. These results underscore the utility of our approach in nominating druggable genes of relevance to psychiatric disorders. Genes supported by multiple independent analytical layers likely represent robust candidates for further studies and therapeutic exploration.

Discussion

In this study, we used a population-level single-cell approach to integrate principles of network medicine with analyses focused on elucidating the genetic basis of psychiatric disorders. In contrast to previous efforts that used bulk transcriptomics data for drug repurposing, our study leverages high-resolution brain-cell-type data across three disorders, highlighting drugs relevant for targeting specific cell clusters. Furthermore, matched gene expression and genotype information from the same donors enabled eQTL analysis of drug target genes, paving the path for applications in precision medicine. Our approach allowed us to expand the druggable genome by focusing on TF regulons within specific cell types. Using a combination of unsupervised (co-regulated gene modules) and supervised (GNN) machine learning approaches, we identified novel genes associated with psychiatric disorders. These genes act as a lever for a more nuanced analysis of repurposable drugs at the cell-type level, a strategy previously not possible due to a lack of sufficient datasets with replicate cohorts.

Our approach found several TFs whose disease expression signature can be reversed by certain drugs. A probable mechanism for drugs indirectly modulating TFs could be by affecting upstream signaling pathways, epigenetic states, or intermediary proteins that ultimately influence a TF’s network activity. Our findings highlight TFs with increased network influence in the disease state, as measured by positive DI scores derived from network bottleneck metrics. We found several TFs with high DI scores whose expression can be reversed by certain drugs, as shown in Figures 2E, 2F, and S4D. These TFs may represent key regulatory hubs involved in transcriptional reprogramming underlying the disease. If such TFs function as activators of disease-associated expression signatures, they could serve as candidates for therapeutic targets. However, the regulatory landscape is complex, and TFs can be involved in pathogenic and protective mechanisms. Further functional validation is necessary, including perturbation experiments to test whether inhibition of high-DI TFs results in reversal of disease-associated expression signatures.

Our study details an approach for expanding network-based drug discovery, previously limited to cell-type-naive protein-protein interaction (PPI) networks, to single-cell genomics within cellular contexts. This allows us to identify drugs that can target specific cell types. For instance, we found a large number of drugs for excitatory neurons in BPD, whereas inhibitory neurons seem to be more amenable to treating SCZ (Figure 5B). The prioritization of a large number of serotonin uptake inhibitors, typically used to treat depression, indicates that our approach mostly identifies compounds that are designed to treat symptoms common across psychiatric disorders. This also suggests the development of more refined models that can pinpoint drug combinations that better suit the peculiarities of individual disorders.

While mimicker and reverser drug relationships suggest potential therapeutic targets, it is important to note that these observed expression changes might reflect compensatory mechanisms rather than primary drivers of pathology. As a result, not all expression-based reversals will necessarily translate into disease-modifying effects. Thus, the repurposable drugs nominated by our study are a first-pass, network-based prioritization strategy. Further experiments are needed to estimate these relationships at different dosages and time points for better clinical translation. While most drugs are systemically administered and may affect multiple cell types, our approach identifies the cell types most likely to mediate therapeutic effects based on disease-specific expression and drug perturbation profiles. This prioritization can inform future development of targeted delivery systems. Thus, our study offers a valuable framework for cell-aware drug repurposing in psychiatric disorders.

The drug-cell eQTL approach allowed us to identify genomic variants that exhibit strong associations with variability in the expression of drug target genes. To the extent that patient-specific drug response is influenced by the underlying expression of a given drug’s target genes, these drug-cell eQTLs identify specific genomic loci that may influence drug response in genotype-specific ways by modulating expression at the level of distinct cell types. Thus, our analysis provides a framework to better understand how genotype can impact drug response at the cellular level, thereby offering new insights into precision medicine-based strategies. However, we note that our framework should not be interpreted as offering drug-response QTLs in a direct sense, as that would entail using drug response data from the same group of individuals for which the genotypes are available, which currently remains limited in publicly available datasets. Nevertheless, the drug-cell eQTLs identified can potentially offer valuable guides for patient stratification in clinical trials. For instance, our results list drugs that might be more susceptible to response variation among individuals (drugs with a large number of associated SNPs).

Limitations of the study

It should be noted that, while the LINCS dataset provides valuable insights into transcriptional perturbations, its ability to accurately predict drug repurposing efficacy remains limited by the use of immortalized cell lines and short-term gene expression measurements. Furthermore, a large fraction of the LINCS L1000 dataset we used is imputed and limited to neuronal cells from the CNS. Thus, the reverser and mimicker effects we identify would need to be further refined with observations in multiple cell lines of the human brain. Another important limitation of our study is the lack of consideration for dose dependency in drug-gene interactions. The LINCS dataset limited our estimations of mimicker-reverser relationships to a 10 mM concentration. Future extensions of this framework could incorporate dose-response information, improving the biological and clinical relevance of predicted candidates.

Our drug-cell eQTL framework does not integrate any pharmacodynamic or pharmacokinetic response data collected from the same individuals whose genotypes were profiled. Thus, we were unable to assess how genetic variation modulates drug responses at the individual level. This precludes the direct inference of genotype-dependent drug efficacy and toxicity. Future extension of our work will require paired genotype and clinical drug response data to highlight genetic variations that influence drug responses. Additionally, comparing our findings with real-world drug performance, including electronic health records, clinical trial data, animal models, and patient cohorts, will be essential to validate our current predictions.

Resource availability

Lead contact

Requests for further information, resources, and software should be directed to and will be fulfilled by the lead contact, Daifeng Wang (daifeng.wang@wisc.edu).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •This paper analyzes existing, publicly available data. The links for the datasets are listed in the key resources table.
• •All original code has been deposited at github.com/daifengwanglab/PEC2NetMed, and the data are publicly available under https://doi.org/10.5281/zenodo.14252962. All derived data from this study can be accessed at https://phase3.gersteinlab.org/netmeds/.
• •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgments

We acknowledge the National Institutes of Health grants RF1MH128695 (to D.W.), U24MH136793 (to M.G.), R21NS127432 (to D.W.), R21NS128761 (to D.W.), U01MH116492 (to M.G.), and P50HD105353 (to Waisman Center); the National Science Foundation Career Award 2144475 (to D.W.); the Simons Foundation Autism Research Initiative pilot grant 971316 (to D.W.); and the start-up funding for D.W. from the Office of the Vice Chancellor for Research and Graduate Education at the University of Wisconsin-Madison. The funders had no role in study design, data collection and analysis, decision to publish, or manuscript preparation. We would like to express our deep gratitude to the patients and their families who generously donated the invaluable biological material essential for the success of this study. We are profoundly indebted to their participation and commitment to advancing scientific knowledge and improving human health.

Author contributions

All individually named authors contributed substantially to the paper through either data generation or analysis: data generation, K.H.A., S.K., X.Z., and Y.X.; and data analysis, C.G., N.C.K., D.C., J.J.C., and C.D. The following two corresponding authors co-led the analysis: D.W. and M.G.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Deposited data
Disorder cell type GRNs	This paper	https://zenodo.org/records/14252962; https://phase3.gersteinlab.org/netmeds.html
snRNA-seq	Emani et al.8	https://brainscope.gersteinlab.org/key_resource_files.html
scATAC-seq	Emani et al.8	https://brainscope.gersteinlab.org/key_resource_files.html
LINCS	Evangelista et al.59	https://maayanlab.cloud/sigcom-lincs/#/Download
B3DB	Meng et al.69	https://github.com/theochem/B3DB
DGIDB	Cannon et al.70	https://dgidb.org/
Software and algorithms
Drug repurposing pipeline	This paper	Github: https://github.com/daifengwanglab/PEC2NetMed/ Zenodo: https://doi.org/10.5281/zenodo.16756007
SCENIC	Aibar et al.71	https://scenic.aertslab.org/
Seurat	Wolf et al.72	N/A
GRNBoost2	Moerman et al.73	https://arboreto.readthedocs.io/en/latest/
MAGMA	De Leeuw et al.44	https://cncr.nl/research/magma/
QTLtools	Delaneau et al.74	https://qtltools.github.io/qtltools/
WGCNA	Langfelder and Horvath75	https://cran.r-project.org/web/packages/WGCNA/refman/WGCNA.html

Method details

The main goal of our study was to develop and apply a network-based drug-repurposing pipeline that leverages single cell multi-omics data for psychiatric disorders. As illustrated in Figure S1, our pipeline integrates multiple steps to systematically identify drug candidates for psychiatric disorders (Figure S1). We began by constructing cell-type-specific gene regulatory networks (GRNs), integrating proximal TF–target relationships inferred from snRNA-seq data with distal regulatory links derived from scATAC-seq data (described in detail below). These GRNs serve as the foundation for identifying cell-type-specific disease-relevant genes. Given that publicly available disorder gene sets are largely cell type–agnostic, we employed a graph neural network (GNN) framework to learn the topological features of known disorder genes within each GRN. The GNN was trained to recognize network patterns associated with these genes and subsequently predict additional candidate genes likely involved in the disorder, within specific cell types. Using the predicted disorder genes, we then applied a network proximity-based drug repurposing strategy to identify drugs that could potentially reverse disease-associated transcriptional signatures. Finally, we explored eQTLs that may modulate the expression of the predicted drug targets, offering insights into personalized medicine.

Cell-type gene regulatory network construction for drug repurposing in psychiatric disorders

snRNA-seq and scATAC-seq data processing

The psychENCODE2 (PEC2) dataset encompassing snRNA-seq and scATAC-seq samples from 388 donors across multiple cohorts has been uniformly processed as described in the recent publication⁸. We used the pre-processed data, including cell clustering and annotation schemes used in the original publication, for the analysis presented here. Figure S1 depicts the flow of data through our network-based drug repurposing pipeline (Figure S1). We constructed two sets of disorder and control GRNs for each cell type. To be consistent in cross-cohort and cross-cell-type comparisons, we built GRNs only for cell types that contained at least 100 cells pooled across donors within each cohort. The first set of GRNs are built for only major cell type clusters, namely, excitatory and inhibitory neurons, astrocytes, oligodendrocytes, oligodendrocyte progenitor cells (OPC), microglia, vlmc, and endothelial cells. Furthermore, for each disorder, we build control and disorder GRNs for each cell type separately. This split of the dataset allowed us to make cross-cohort contrasts and comparisons, as well as test for reproducibility for disorders with redundant cohorts. We built the second set of GRNs for subclass level cell types which include ‘L2.3.IT’, ‘L4.IT’, ‘L5.IT’, ‘L5.ET’, ‘L5.6.NP’, ‘L6b’, ‘L6.IT’, ‘L6.CT’, and ‘L6.IT.Car3’ subclasses of excitatory neurons, and ‘Lamp5’, ‘Pvalb’, ‘Sncg’, ‘Sst’, ‘Lamp5.Lhx6’, ‘Vip’, ‘Pax6’, and ‘Chandelier’ subclasses of inhibitory neurons. The set 2 cell type GRNs were built for each disorder separately, combining redundant cohorts for SCZ. Altogether, cells from a total of 247 donors (36% females) were used for GRN inference. The data subset comprises a total of 1,64,5813 cells, with 901,101 (54.76%) disorder cells and 744,712 (45.24%) control cells (Figure S2). The following procedure was applied to predict the cell type by disorder GRNs.

Gene regulatory networks

We first normalized raw counts within a cell cluster using the ‘NormalizeData’ function of Seurat v5⁷⁶. We then filtered genes with low variation in expression across a cell cluster in a given cohort using Seurat’s ‘FindVariableFeatures’⁷² and created a list of top 5000 highly variable genes (HVG). To the HVG list, we added TFs for which binding sites (TFBS) are known⁷⁷. We reason that because TFs are regulatory genes, subtle changes in their expression patterns can have profound downstream effects. Excluding such TFs from analysis may collapse the GRN making it sparse. Subsequently, we supplied the expression matrix consisting of appended HVGs to the GRNboost2 algorithm⁷³, along with the list of TFs, to predict their targets. Subsequently, we provided the output of GRNBoost2 to SCENIC to filter TF→TG links that are not representative of TFBS within gene promoters⁷¹. Briefly, SCENIC uses cisTarget to perform motif enrichment analysis by scanning the promoters of coexpressed genes to identify overrepresented TF binding motifs. Only transcription factors whose motifs are statistically enriched are considered potential regulators of the genes. Such TF→TG links were labeled as proximal edges. For distal links, we used enhancer-promoter links from scATAC-seq data and mapped TFBS on these interaction regions, thereby linking TFs to targets via enhancers (as described previously in the original PEC2 publication⁸). We further filtered the links from scATAC-seq data if there was no coexpression between the TF→TG pair in the cell cluster (detected in GRNboost2 runs) and labeled the remaining links as distal edges. This step allowed us to add variation to distal links in sub class GRNs since the scATAC-seq data is only available for the major class cell types. Finally, for each cell type disorder combination, we merged the proximal and distal edges into composite GRNs made available for further analyses.

GRN evaluation

To assess the robustness of the inferred GRNs, we examined the sensitivity of our pipeline to sampling biases. We grouped the TF→TG links into ten deciles based on decreasing edge importance scores as inferred by SCENIC. We then calculated the percentage of edge overlap in each decile between SCZ GRNs from the CMC and SZBD cohorts for each cell type, repeating the process for control GRNs as well. Our analysis revealed a relatively higher overlap in the top deciles, with a gradual decline and minimal overlap in the 10th decile (Figure S3A). This suggests that high-confidence GRN links are reproducible, while lower-confidence links may be more sensitive to sampling variations. We validated our predicted TF–target links against ChIP-seq data from the Gene Transcription Regulation Database, which contains uniformly processed TF binding sites from public datasets. Since peaks from different conditions are pooled, the GTRD lacks cell type information but remains useful for assessing TF motif presence near predicted targets. We calculated the overlap between predicted TF–target pairs and GTRD peaks (±5 kb from TSS) and found an average precision of over 62% across cell types and disorders (Figure S3B). The lower recall is expected due to the lack of condition-specific data in the GTRD.

Druggable cell-type transcription factors for psychiatric disorders

Hubs, bottlenecks, and disorder influence score

We counted the number of outgoing edges from a given TF in a given GRN as its outdegree. We used betweenness centrality to quantify the number of times a TF appears in the shortest paths between pairs of target genes. To assess the influence of TFs in disease contexts, we calculated the Disease Influence (DI) score using betweenness centrality as a proxy. The DI score of a TF is defined as the log2 ratio of the betweenness centrality in disease gene regulatory networks (GRNs) to the betweenness centrality in control GRNs (Figure S5).

$$DI=\left(\frac{B{C}_{disease}}{B{C}_{control}}\right)$$

Where $B{C}_{disease}andB{C}_{control}$ are betweenness centrality in disease and control GRNs, respectively.

TF specificities

We profiled TF specificity using betweenness centrality and outdegree for diseases and cell types. We selected TFs based on their top 10% outdegrees and betweenness centralities, respectively. We then counted the occurrence of these TFs across diseases and cell types and filtered out TFs that appeared in only one disease or cell type (referred to as disease-specific TFs and cell type-specific TFs, respectively). Finally, we identified TFs that are specific to diseases and to cell types based on outdegree, and separately identified TFs that are specific to diseases and to cell types based on betweenness centrality. We categorized these TFs into disease-specific and cell type-specific groups based on betweenness centrality and outdegree, respectively (Figures S4A and S4B).

Druggable regulons

To identify druggable regulons, we obtained gene expression profiles from chemically perturbed human cell lines available through the SigCom LINCS website⁷⁸. Specifically, we worked with the Characteristic Direction (CD)-processed L1000 signatures. The CD method is a computational approach used to analyze gene expression data within the Library of Integrated Network-based Cellular Signatures (LINCS) L1000 dataset. Its primary objective is to pinpoint the most significant gene expression changes in response to various perturbations. The CD method claims to offer a robust ranking of differentially expressed genes, based on their contribution to the overall perturbation effect.

Next, we obtained a list of drugs predicted to permeate the blood-brain barrier (BBB) from the B3DB database⁶⁹ and intersected these BBB-permeable drugs with those in the LINCS L1000 CD dataset. For the matching drugs, we compared each TF’s expression in response to chemical perturbations in the LINCS dataset to the gene expression observed in the respective disorders. We found several TFs with high DI scores whose expression can be reversed by certain drugs, as shown in Figures 2E, 2F, and S4D.

Disorder risk genes converge on coregulated gene modules

Coregulation modules

Our main goal was to prioritize repurposable drugs for the three psychiatric disorders in a cell-type specific manner using a network based approach. In our initial survey, we found only one TF (PPARG) listed as a direct drug target. This indicated that most drug targets are non-TF genes. Therefore, it is imperative to GRNs from the perspective of non-TF, or target genes (TG). To align with the GRNs presented in this study, we converted the TF→ TG edges into TG→ TG edges using the following procedure.

First, we used the Jaccard’s Index (JI) to quantify the overlap between the regulators of a TG pair in a GRN. We then arranged the resulting JI values as a dissimilarity matrix (1 - JI) with TGs in rows and columns. The dissimilarity matrix was then supplied to a hierarchical clustering algorithm using the flashClust function of WGCNA⁷⁵. The dendrograms of the resulting clustering were then cut using cutreeDynamic function (with parameters method="hybrid", deepSplit = 2, pamRespectsDendro = FALSE, msize = 30). The msize = 30 parameter ensures a minimum module size of 30 for a respectable sample size for downstream enrichment analysis as described below. Note that the scATAC-seq data is availabe for only the major cell type classes. Therefore, we used only proximal links for identification of modules to retain cell type-and-disorder specificity in predicted modules. Shared distal links between control and disorder, and between subclasses of neuronal cell types might obscure biological signals for downstream analysis. Note that we used only proximal links for identification of modules because the shared distal links between control and disorder might obscure biological signals.

Magma enrichment

We used the MAGMA (Multi-marker Analysis of GenoMic Annotation) to perform LD aware enrichment of GWAS traits within module gene sets and other gene sets described in the main text. To do this, first we obtained a number of GWAS results and extracted summary statistics of SNPs along with nominal p-values (p <=0.05). The SNPs were then annotated by applying the annotate function using gene locations from the human genome build 38 obtained from NCBI. The reference files used for enrichment were based on European ancestry in Phase 3 of the 1,000 Genomes project. The sample size parameter N was set according to the sample size mentioned in the corresponding GWAS publications as follows: SCZ: 76755, ASD: 46350, BPD: 41917, epilepsy: 29944. The resulting p-values from MAGMA runs are shown in the main Figure 3B.

Gene ontology enrichment

To test the enrichment of gene sets from the GO catalog, we downloaded the GMT formatted GO term annotations from the Enrichment Map database available (https://download.baderlab.org/EM_Genesets/). The GMT files were read into the R using the GSA library (version 1.03.2). Then, biological process terms with more than 300 or less than 30 genes were filtered and the remaining terms used to calculate the over-representation of annotated genes annotated within each of the query gene sets (e.g. modules). The statistical significance of the overlap was calculated using a hypergeometric test using the sum of all genes across all modules as the background universe. The resulting p values were corrected for multiple testing using the Benjamini-Hochberg procedure and enrichments with FDR <= 0.1 are reported. These operations were performed in R using the HypeR package in R.

Risk gene enrichment

For each disorder, a set of known risk genes were extracted from public databases as follows. For ASD, we obtained the Gene Scoring module from the SFARI Gene database. This module offers a comprehensive scoring indicating the strength of literature evidence supporting each gene's association with autism. We filtered the list to retain only genes labeled as ‘Category 1’ and ‘Syndromic’, as high confidence autism risk genes. For BPD and SCZ, we obtained disease gene associations listed in the DisGeNET database. For both disorders, we retained only those genes that are labeled as “CTD_human” and used them for enrichment analysis.

Enrichment of differentially expressed genes

To calculate the enrichment of differentially expressed genes within modules, we obtained the summary statistics for whole cortex differential gene expression results from ASD versus control from a previously published paper. The log fold change values were used as a parameter. Briefly, the parametric analysis of gene set enrichment (PAGE) is a statistical method used to determine whether predefined sets of genes show statistically significant, coordinated differences in expression in different biological conditions or treatments. Unlike non-parametric methods like GSEA, which rank genes based on their expression levels and use permutation testing, PAGE employs a parametric approach that relies on the normal distribution to model gene expression changes. Thus, the PAGE approach provides higher statistical power and sensitivity when detecting coordinated shifts in expression across entire modules, rather than identifying enrichment driven by a subset of top-ranked genes. Under this framework, we computed a Z-score for each gene set (module) based on the mean fold change of the genes within the set. The Z-score indicates how far the mean of the gene set is from the overall mean expression changes in units of standard deviation.

$$Z=\frac{\underset{\_}{d}-{\mu }_d}{{\sigma }_d}$$

where $\underset{\_}{d}$ is the mean expression change of the module, ${\mu }_d$ is the overall mean of all genes, and ${\sigma }_d$ is the standard deviation across all fold change values. The resulting Z-scores are shown in main text Figure 3.

Graph neural networks predict novel cell-type disorder genes

Label creation for supervised learning

We created two sets of training labels for disorder cell type gene prioritization models for the purposes of comparison. For model 1, we used known disorder risk genes from public databases as positive labels. Specifically, for ASD, we used candidate gene nominations within category gene.score = 1 or syndromic = 1 in the SFARI database. For SCZ and BPD, we selected genes within the DisGeNet database labeled as “CTD_human”. Note that these sets are devoid of cell type specificity. For model 2, we appended model 1 labels with genes in modules enriched with known risk genes, thereby leveraging risk neighborhoods of the networks.

Model training and evaluation

The model consists of a standard graph convolutional network (GCN) with 2 hidden layers and an output of dimension 2 for $n$ genes. Batch normalization layers are included after each hidden layer and immediately followed by Leaky ReLU activation functions. Typically, the dimension of the first and second hidden layers are set to 32 and 16, respectively. We refer to the first three layers as EMB and the last as CLASS. As we use no node features, the input, $X_i$ for each gene $i$ is equal to a zero-array of length $n$ with $1$ at position $i$. Each model has an additional FCL to predict the weight of a directed edge given embeddings corresponding to each gene. The models each use two losses: (1) Embedding loss, which is the MSE of the predicted adjacency matrix by the model compared to the actual adjacency matrix and (2) classification loss, which is the cross entropy loss between the model output layer and the true labels. These may also be represented as

$$l_{Embedding}=\frac{1}{n^2}\sum _{i=1}^n\sum _{j=1,j\ne i}^{n}FCL\left(EMB\left(X_i\right),EMB\left(X_j\right)\right)-X_{ij}$$ (Equation 1)

$$\sum _{i=1}^n{Y}_i·\mathit{log}\mathit{log}CLASS\left(GCN\left(X_i\right)\right)+\left(1-Y_i\right)·\mathit{log}\mathit{log}CLASS\left(GCN\left(X_i\right)\right)$$ (Equation 2)

Where $n$ is the number of genes, ${embedding}_i$ is the embedding associated with gene $i\text{,}$ and $X_{ij}$ is the weight of the edge between genes $i$ and $j$. The two losses iterate the model separately, with the embedding layers frozen for the backpropagation of the classification loss.

Each cell type is trained individually using five fold cross-validation. More precisely, the data is split into 5 partitions (folds) and a model is trained using each partition as validation data, resulting in 5 separate models. The mean and standard deviation of the scores for each gene are then recorded and taken as the final model predictions. The folds are analyzed individually for performance statistics such as AUROC and AUPRC.

BBB Drug prioritization

For each cell type, we examine the top decile of genes prioritized within the model. Then, the resulting gene list is scanned for gene targets of drugs which were earlier predicted to cross the blood brain barrier using ML models. These counts are then shown in Figure 4C for both models 1 and 2.

Network-based drug repurposing

BBB permeable drug-target network

We acquired an academic license for the DrugBank database and extracted all drug IDs along with their corresponding target genes. Additionally, we obtained data from the Drug-Gene Interaction Database (DGIDB), from which we filtered out compounds classified as "approved == TRUE" and "anti_neoplastic == FALSE." Concurrently, we obtained blood-brain-barrier (BBB) permeability labels for small molecules from the B3DB. The B3DB contains includes both numerical (log BB) and categorical (BBB+ or BBB−) data for over 7,800 compounds, sourced from over 50 publications. It contains both quantitative (log BB values) and qualitative (BBB+ or BBB−) measurements for drug compounds. The database is designed to aid machine learning models for predicting BBB permeability, addressing limitations of smaller, less diverse datasets and making it a valuable resource for drug discovery targeting central nervous system diseases. We matched the IUPAC names of BBB+ molecules with those listed in DrugBank to extract the drug IDs for brain disorders. We cross-referenced the names of BBB+ molecules with the drug_claim_name entries in DGIDB and combined them with the BBB+ compounds from DrugBank to create a network of 802 drugs along with their 1196 target genes for further analysis, as described below.

Network proximity

We leverage the relationships between drugs, their targets, and disease-associated genes within a cell type network for drug repurposing. This approach to identify potential drug candidates that can be repurposed to treat diseases based on their "proximity" to disorder-associated nodes within a network graph. To create a pool of disease/disorder specific nodes, we utilized novel predictions from our GNN models along with the positive training labels. This allows us to use cell type relevant risk genes to augment current knowledge on the genes underlying the disorder. Further, we used cell type coregulatory networks as graphs for drug repurposing instead of using bipartite GRNs to avoid degree and connectivity biases.

To analyze the network proximity between drug targets and disorder genes in a coregulatory network, we compute various statistics to assess whether drug targets are closer to disorder genes than would be expected by chance. First, we calculate the degree distribution of both drug targets and disorder genes, which represents the number of connections each node has in the network. This is important for ensuring that our random selection of nodes for baseline comparison maintains similar connectivity properties. Next, we calculate the proximity between two sets of nodes (drug targets and disorder genes) based on their shortest path distances in the network. We then compute the mean of the minimum distances between each pair of nodes from drug targets to disorder genes. For a set of drug targets T and a set of predicted risk genes G, the proximity score dTG is calculated²⁵:

$$d_{TG}=\frac{1}{\left|T\right|}\sum _{t\in T}\underset{g\in G}{\min }d\left(t,g\right)$$

where d(t,g) denotes the shortest path length between nodes t and g in the network.

To create a baseline for comparison, we compute a random distribution of proximity values by randomly selecting nodes from the network that match the degree distribution of the actual drug targets and disorder genes. We calculate their proximity iteratively 100 times and use this random distribution as a baseline to assess the statistical significance of the observed proximity.

We compute a Z-score of the observed proximity as:

$$Z=\frac{d_{TG}-{\mu }_{rand}}{{\sigma }_{rand}}$$

where μ_rand and σ_rand are the mean and standard deviation of the null distribution of proximity scores. A significantly negative Z-score indicates that drug targets are closer to the risk genes than expected by chance. P-values were computed by converting Z-scores to cumulative probabilities under the standard normal distribution using the lower tail (i.e., p = P ( ≤ z ). All observations with a p-value < 0.1 are reported. All these operations were performed using the igraph library in R (https://igraph.org).

Reversers vs mimickers

We obtained the Library of Integrated Network-Based Cellular Signatures (LINCS) dataset from SigCom. Specifically, we worked with the L1000 Chemical Perturbations dataset that provides up and down regulated gene sets in response to chemical perturbations measured across several human cell lines. We extracted experiments conducted in CNS cell lines (“NEU” cell lines) with the repurposable drug molecules identified in our study. In parallel, we obtained results from the differential gene expression (DGE) analysis for the disorders reported by us previously⁸. We tested the enrichment of drug-induced up- and- down-regulated genesets in the L1000 data in the disorder DGE data. To assess statistical significance of enrichment, we performed a two-sample Kolmogorov–Smirnov (KS) test comparing the distribution of log₂ fold changes for drug-regulated genes against that of all other genes within the same cell type. We corrected the resulting KS-test p-values for multiple correction using the Benjamini-Hochberg procedure and selected drugs with adjusted p-values <= 0.05. We classified a drug as a ‘mimicker’ if the drug's upregulated gene set showed a positive mean fold change in disease, and as a reverser otherwise. Conversely, we if a drug's downregulated gene set showed a negative mean fold change, we classified it as a mimicker; otherwise, as a reverser. We reported only drugs that showed consistent classification as either mimicker or reverser across both their ‘up’ and ‘down’ gene sets within the same cell type and disorder. This framework enables the identification of drug candidates that robustly mirror or counteract disease-associated gene expression patterns in a cell-type-specific and interpretable manner. Then, for the set of up- and- down-regulated genes for each repurposable drug in the L1000 data, we calculated the mean fold change in disorder. We then compared the direction of changes and labeled drugs as ‘reverser’ if the mean is in the opposite direction to that reported in the LINCS1000 dataset for both up- and- down- sets for the drug, or ‘mimiker’ otherwise.

Sensitivity analysis

We wanted to test the sensitivity of drug repurposing to edge-threshold parameter. To do this, first, we evaluated the robustness of TF hierarchy, since the hierarchical levels in GRNs reflect the regulatory influence of TFs and may affect downstream gene expression dynamics. We randomly downsampled 90% to 50% of the nodes/edges within the TF subnetworks of each cell-type-specific GRN. Then, we determined the hierarchy levels of the TFs using the downsampled TF subnetworks and compared these new hierarchy levels to those observed in the original TF subnetworks (Table S12). We also subsetted the GRNs to select only top 75% and top 50% edges, sorted by the edge importance scores. We then reran the drug repurposing pipeline for a few randomly picked cell types and compared the overlaps of nominated drugs between each subsetted GRN and the full GRN. In addition to the GRNboost2 for GRN inference, we used a mutual information (MI) as an alternate method to infer GRNs for a few randomly picked cell types. The MI-based GRN inference is a popular strategy extensively used previously.^79,80 We processed the MI-based though SCENIC’s motif enrichment pipeline and the resulting GRNs were used for comparison with the GRNboost2-based GRNs. We used the scikit-learn’s implemented MI function to estimate the MI between TF-target gene pairs. (https://scikit-learn.org/stable/modules/generated/sklearn.feature_selection.mutual_info_regression.html).

Drug-cell eQTLs

We carried out a set of QTL analyses to identify cases for which genotypic variation is associated with significant changes in expression for the gene sets targeted by different drugs. This entailed using a total sample size of 125 (after filtering out duplicate samples or samples with missing covariates, and by including only those samples for which genotype PCs had been calculated⁸. As the response variable in this QTL analysis, for a given drug found to be repurposable in our analysis, we first identified the set of genes which are known to be affected by that drug. For a given drug, we refer to this set of genes as the drug’s “drug target genes”. Then, for each donor, we log-normalized the gene expression matrix and calculated the average expression of the drug’s drug target genes within each cell type. These average expression values, for every drug-cell type combination, were then used to populate a ‘DrugCell’ matrix (here, the term ‘DrugCell’ is adopted as a short term for a drug-cell type combination). In this matrix, rows represented drug-cell combinations, and columns corresponded to donors. The DrugCell matrix was then used as input to eQTL analysis.

We used QTLtools (version 1.3.1)⁷⁴ and ran our calculations in nominal-pass mode. Because each DrugCell comprises multiple genes, the nominal-pass calculation was performed by placing each DrugCell in each of the 22 autosomal human chromosomes, and then performing a search for SNPs within a cis-window that engulfs each entire chromosome in full. We used biological sex, age of death, diagnosis, cohort, five genotype PCs (to control for ancestry), and 5 expression PCs (to control for hidden batch effects) as covariates in our analysis. We filtered out low-frequency variants by removing any variant with a minor allele frequency (MAF) of less than 0.05. This yielded a total of 1,953,349 SNPs.

Specifically, we used the following format as the command when running QTLtools (again, the large cis window was used to ensure that our calculations covered the entirety of each chromosome):

QTLtools cis --vcf vcf_file_containing_genotype_variants --bed bed_file_containing_drugcell_phenotype_values --cov matrix_containing_covariates --nominal 2.5597064323886822e-08 --window 300000000 --normal --out output_file_name

Thus, in sum, this analysis is carried out by performing linear regression, wherein we included covariates as well as the response variable (i.e., DrugCell values) in a linear regression on genotype dosage. The residual DrugCell values were rank normal transformed (after the covariate correction). In order to control for the extremely large number of tests for each individual DrugCell (namely, the number of filtered SNPs), P-values in this analysis were corrected using Bonferroni correction at the level of 0.05, with this correction having been applied at the level of each DrugCell separately (thus, the nominal p-value threshold was ∼2.56e-08 = 0.05/1,953,349). This yielded a total of 335 significant Drug-associated cell-type eQTLs, and we make these available as Table S10. The format for this output file is given as follows:

• (1)The drugCell name
• (2)The variant ID
• (3)The variant chromosome
• (4)The start position of the variant
• (5)The nominal p-value of the association between the variant and the DrugCell
• (6)The r2 of the linear regression
• (7)The beta (slope) of the linear regression

We performed a global survey using all genes (targets of drugs) via the Gene Ontology enrichment. We compiled a comprehensive list of the drugs that comprise the significant drug-cell eQTLs. Using this set of “significant drugs”, we constructed the full set of genes that are targets of these drugs. The GO enrichment analysis was then performed on this set of 56 drug target genes. We used g:Profiler⁵⁰ to perform GO enrichment analysis for molecular functions and biological processes. For ease of visualization, we used an adjusted p-value cutoff of 10-7 to generate the GO enrichment plot show within Figure S13.

Gene ranking and prioritization

To integrate various lines of evidence, we worked with targets of drugs with significant mimicker-reverser relationships found in our study. We evaluated each gene target across four distinct analytical layers, including 1) differential expression status from psychENCODE2 single-nucleus RNA-seq data, 2) gene-level MAGMA analysis assessing enrichment of GWAS signals within disease-associated modules, 3) graph neural network (GNN) predictions of candidate risk genes, 4) associations from drug-cell eQTL analyses. Each gene was assigned binary presence/absence calls across these layers, resulting in a total of 48 distinct lines of evidence. We computed a cumulative evidence score for each gene based on the number of layers supporting its involvement. Genes were then ranked by their total support to identify high-confidence candidates. This list is provided in Table S11.

Development of psyDGN

We developed an R shiny web application, psyDGN (https://daifengwanglab.shinyapps.io/psyDGN/), allowing users to interactively explore the findings of the current study. The psyDGN consists of three sub-applications. In the first sub-application, it allows users to visualize cell-type gene regulatory networks (GRNs). Users can select one of the three disorders (autism spectrum disorder, bipolar disorder, and Schizophrenia) and a cell type to visualize a corresponding disease-cell-type GRN. Users can also filter GRNs based on the edge importance scores and isolate connections specific to selected transcription factors. In the second sub-application, users can tabulate the predicted disorder genes for a given cell type and a disorder. In the third sub-application, users can visualize cell type GRN overlaps with drug targets. The psyDGN allows users to select a drug and a cell type of their choice from a drop-down menu and outputs three networks corresponding to the three disorders. Only the transcription factor-target gene connections overlapped with the drug-target genes are visualized.

Published: September 18, 2025